xphirobotics/22hr_data

XP Robotics Egocentric Activities (22h)

A multimodal egocentric dataset of everyday manual activities, collected and released by XP Robotics. Each recording captures a first-person view of a person performing a task, together with synchronized motion, position, and audio signals.

Overview

• 152 recordings, 23.7 hours total, across 6 activity categories
• Single egocentric camera, AV1-encoded video
• Synchronized IMU, GPS, audio, camera pose, and coordinate transforms
• Stored as MCAP with Foxglove/protobuf message schemas

Category	Recordings	Duration (min)	Size (GB)
Assembling furniture	24	218	23.1
Commercial kitchen activity	41	272	29.3
Knitting/Sewing	39	522	56.7
Tidying and organizing	14	139	14.9
Vehicle maintenance	21	160	16.6
Yard care	13	111	12.5

Recording schema

Every recording is a single MCAP file containing the following synchronized streams:

Topic	Schema	Description
/top-left-camera/image-raw	foxglove.CompressedVideo	Egocentric camera video (AV1)
/top-left-camera/camera-info	foxglove.CameraCalibration	Camera intrinsics / calibration
/camera/pose	foxglove.PoseInFrame	Camera pose over time
/tf	foxglove.FrameTransforms	Dynamic coordinate transforms
/tf-static	foxglove.FrameTransforms	Static coordinate transforms
/imu/data	iw.Imu	Filtered IMU (orientation, angular velocity, acceleration)
/imu/data_raw	iw.Imu	Raw IMU measurements
/gps/fix	foxglove.LocationFix	GPS location fix
/audio	foxglove.RawAudio	Audio stream
/metadata/header	iw.HeaderMetadata	Per-recording metadata (resolution, frame rate, codec, task)
/metadata/footer	iw.FooterMetadata	Per-frame image timestamps

Technical details

• Video: AV1, YUV420P, ~30 fps. Resolution is 1920x1080 (125 recordings) or 1920x1440 (27 recordings).
• Camera: single egocentric view, plumb_bob distortion model, intrinsics in /top-left-camera/camera-info.
• Motion: 6-axis IMU at the camera frame rate, both filtered (/imu/data) and raw (/imu/data_raw).
• Position: GPS fixes in /gps/fix.
• Audio: raw audio in /audio.

Directory structure

22hr_data/
  metadata.csv
  assembling_furniture/        assembling_furniture_001.mcap ...
  commercial_kitchen_activity/ commercial_kitchen_activity_001.mcap ...
  knitting_sewing/             knitting_sewing_001.mcap ...
  tidying_and_organizing/      tidying_and_organizing_001.mcap ...
  vehicle_maintenance/         vehicle_maintenance_001.mcap ...
  yard_care/                   yard_care_001.mcap ...

Files follow a uniform <category>/<category>_<NNN>.mcap naming scheme, numbered by recording start time within each category. metadata.csv lists every file with its category, duration, size, resolution, video format, start time, and message count.

Reading the data

The recordings use protobuf message encoding. Read them with the mcap and mcap-protobuf-support packages:

from mcap.reader import make_reader
from mcap_protobuf.decoder import DecoderFactory

with open("knitting_sewing/knitting_sewing_001.mcap", "rb") as f:
    reader = make_reader(f, decoder_factories=[DecoderFactory()])
    for schema, channel, message, proto in reader.iter_decoded_messages(
        topics=["/top-left-camera/image-raw", "/imu/data"]
    ):
        print(channel.topic, message.log_time, type(proto).__name__)

Install with:

pip install mcap mcap-protobuf-support

The files can also be opened directly in Foxglove for visual inspection.

Attribution

Collected and released by XP Robotics.